>
> I guess the real question is, why maintain index flags and not simply
> drop the index entry altogether?
>
> A more interesting case would be to have the backend process record
> index tuples that it would invalidate (if committed), then on commit
> send that list to a garbage collection process.
>
> It's still vacuum -- just the reaction time for it would be much
> quicker.
>
This was my original question.
I guess the problem is with MVCC. The row may have gone from your
current view of the table but not from someone elses. I don't (yet)
understand the way it works to say for sure, but I still think it is worth
pursuing further for someone who does know the deep stuff. They seem
to have concluded that it is not worth it however.
Cheers,
Gary.